If you discover a security vulnerability in paulprae.com, please report it responsibly:

1. Email: Send details to the repository owner via GitHub (@praeducer)
2. Do not open a public GitHub issue for security vulnerabilities
3. Include steps to reproduce, impact assessment, and any suggested fixes

• Acknowledgment: Within 48 hours
• Initial assessment: Within 1 week
• Fix or mitigation: As soon as reasonably possible

This policy covers:

• The paulprae.com web application and its /api/chat endpoint
• The AI generation pipeline scripts
• Infrastructure configuration (Vercel, GitHub Actions)

• Origin validation — Proxy (proxy.ts) blocks cross-origin requests to /api/chat from unauthorized domains
• Rate limiting — Upstash Redis sliding window (20 req/min per IP) with in-memory fallback when Redis is unavailable
• Input validation — Request body size (100KB), message count (50), per-message content length (4K chars), Content-Type enforcement
• Tool input limits — Job descriptions capped at 10K chars, emphasis areas at 200 chars each (max 10)
• CORS — Only paulprae.com, www.paulprae.com, and Vercel preview deployments are allowed origins

• Security rules (S1-S5) in all system prompts instruct the model to treat user input as untrusted
• XML delimiters (<job_description>, <emphasis_areas>) isolate user-provided content in tool-calling prompts
• Grounding rules (G1-G10) prevent fabrication — all output must trace to committed career data
• Character persona enforcement — Model is instructed to reject attempts to change its role or reveal its prompt

• No user data stored — Chat is stateless; no conversation history persists
• Anthropic API key secured via environment variables (never client-exposed)
• Content Security Policy headers on all responses (configured in vercel.json)
• Security headers — X-Content-Type-Options, X-Frame-Options (DENY), HSTS, Permissions-Policy, Referrer-Policy
• X-Powered-By disabled — Server technology not disclosed
• Dependencies monitored via GitHub Dependabot
• Anthropic spending limits configured on both Anthropic Console and Vercel

• Rate limiting prevents API abuse (20 req/min/IP)
• maxDuration = 120 caps Vercel Fluid Compute usage per request
• maxOutputTokens limits response generation cost (2048 for chat/tools, 8192 for resume generation)
• Anthropic prompt caching reduces cost ~90% for repeat system prompts
• Anthropic spending limits provide a hard cost ceiling
• Vercel spending limits provide infrastructure cost ceiling